We will either need to use OpenAI's old Safety Gym, or Safety-Gymnasium, which is maintained by PKU-Alignment.
Safety Gymnasium seems a bit more promising since it is a pip-installable package, but I haven't tried Safety Gym yet (involves a git clone to install).
Safety Gymnasium does have its issues—it seems to be incompatible with Box2D (PyGame specifically throws errors) in the Docker context due to some required package downgrades. This might not be that bad if we're only focusing on MuJoCo and are willing to throw away Box2D environments.
A third alternative is to try to port in the desired car environment directly into our Docker images. This seems really annoying, and I'd like to avoid it if possible. If we can get Safety Gym(nasium) installed, then we get all of the safety-focused environments for free, which seems worth pursuing.
We will either need to use OpenAI's old Safety Gym, or Safety-Gymnasium, which is maintained by PKU-Alignment.
Safety Gymnasium seems a bit more promising since it is a pip-installable package, but I haven't tried Safety Gym yet (involves a git clone to install).
Safety Gymnasium does have its issues—it seems to be incompatible with Box2D (PyGame specifically throws errors) in the Docker context due to some required package downgrades. This might not be that bad if we're only focusing on MuJoCo and are willing to throw away Box2D environments.
A third alternative is to try to port in the desired car environment directly into our Docker images. This seems really annoying, and I'd like to avoid it if possible. If we can get Safety Gym(nasium) installed, then we get all of the safety-focused environments for free, which seems worth pursuing.